Pith. sign in

REVIEW 4 major objections 6 minor 32 references

End-to-end Acoustic-linguistic Emotion and Intent Recognition Enhanced by Semi-supervised Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Semi-supervised learning on 46,542 unlabelled speech samples lifts joint emotion-intent recognition to JRBM 0.430, 0.123 above the acoustic baseline.

desk verdict A solid empirical SSL-for-SER paper whose headline fusion gain is confounded by the lack of a baseline-fusion control. read the letter →

arxiv 2507.07806 v1 pith:ROLMWFEJ submitted 2025-07-10 cs.SD eess.AS

Zhao Ren , Rathi Adarshi Rammohan , Kevin Scheck , Sheng Li , Tanja Schultz This is my paper · ORCID
classification cs.SDeess.AS
keywords speechemotionrecognitionintentsemi-supervisedlearningfix-matchfull-matchmulti-tasklatefusionjointbalancemetrics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Speech emotion and intent recognition usually needs expensive manual labels, but streaming services produce abundant unlabelled speech. This paper asks whether that unlabelled audio can substitute for labels via semi-supervised learning, and answers yes for a 3,610-sample labelled set paired with 46,542 unlabelled samples. The authors compare two pseudo-label schemes, fix-match and full-match, in a multi-task setup where HuBERT processes speech and RoBERTa processes transcripts. Their best result comes from late fusion of the best acoustic and text models: joint harmonic-mean F1 (JRBM) rises to 0.430 on the test split, from 0.307 for HuBERT alone and 0.326 for RoBERTa alone. If this holds, practitioners could build better emotion-and-intent systems using the unlabelled data they already collect.

What carries the argument

The load-bearing machinery is pseudo-label semi-supervision. Fix-match learning generates pseudo labels from weakly augmented unlabelled samples, keeps them only when the model's confidence exceeds a threshold, and trains on strongly augmented versions. Full-match learning additionally applies an adaptive negative loss that suppresses probabilities below the top-k classes and an entropy meaning loss that flattens the middle-ranked classes, consuming all unlabelled samples rather than only confident ones. Both losses are applied separately to the emotion and intent heads of the multi-task model. The acoustic branch uses HuBERT, a masked-prediction self-supervised speech transformer; the text branch uses RoBERTa, a pretrained language model; and final predictions are combined by margin sampling. Performance is measured by JRBM, the harmonic mean of the emotion and intent F1 scores.

What would settle it

Re-run the same HuBERT/RoBERTa recipe on the official MC-EIU test labels (or across five random seeds) and compare JRBM to the 0.307 HuBERT baseline; the central claim fails if the reported 0.430 result does not reproduce.

Watch

Extended reading notes

Core claim

The paper's central claim is that semi-supervised learning with large unlabelled speech and text corpora improves joint emotion and intent recognition over supervised fine-tuning of foundation models. Applying fix-match learning and full-match learning to HuBERT and RoBERTa under multi-task training, the authors report that most semi-supervised configurations beat their respective baselines on both validation and test splits. The strongest configuration is not a single model but a late fusion of the best acoustic and best text models, selected by margin sampling, which reaches a JRBM of 0.430 against 0.307 for HuBERT and 0.326 for RoBERTa. They interpret this as evidence that the two modalities carry complementary information and that unlabelled data can be harnessed without manual annotation.

Load-bearing premise

The whole evaluation leans on the assumption that re-splitting the training set and treating the official validation set as the test set reproduces the challenge's true test conditions, and that the single reported run is representative.

Editorial extensions

If this is right

  • Unlabelled speech collected in the wild can improve both emotion and intent recognition without human annotation, lowering the cost of deploying such systems.
  • Late fusion of acoustic and text models outperforms either modality alone, so a practical system should keep both channels when transcripts are available.
  • Full-match learning helps most when weak augmentation is absent, while fix-match wins with weak augmentation, so the choice of semi-supervised method depends on the augmentation budget.
  • The best two-model fusion beats the best four-model fusion, suggesting that selecting a few complementary models matters more than adding more models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Table II reports one run without error bars or significance tests, so part of the 0.123 JRBM gap could be seed-dependent variance; rerunning with several seeds would tell how much is stable.
  • Because the text semi-supervised branch relies on Whisper transcriptions of unlabelled audio, transcription errors could leak into pseudo-labels; using reference transcripts would isolate the semi-supervised effect.
  • The evaluation uses the official validation set as the test set; confirming on the challenge's true test labels or an independent split would test whether the re-split result transfers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper addresses speech emotion and intent recognition in a multi-task setup using HuBERT for speech and RoBERTa for text, and introduces two semi-supervised learning strategies (fix-match and full-match) that leverage 46,542 unlabelled samples from the MC-EIU dataset. The contributions are the SSL formulation for multi-task emotion/intent classification, a comparison of weak/strong augmentations for unlabelled data, and a late fusion of the best acoustic and linguistic models, reporting a JRBM improvement of 12.3% and 10.4% over the acoustic and text baselines. The evaluation is carried out on a re-split of the MC-EIU Track 1 English data, with the original validation set used as the test set.

Significance. If the reported gains are robust, the paper provides a practical recipe for exploiting unlabelled speech and transcripts in emotion/intent recognition, and its comparison of fix-match and full-match learning under different augmentations is a useful empirical contribution. The use of Whisper-transcribed unlabelled audio for text SSL, the choice of margin-sampling late fusion, and the focus on a joint emotion/intent metric are sensible. However, the evaluation design currently prevents attribution of the headline gains to semi-supervised learning, and the lack of uncertainty estimates leaves the stability of the improvements unestablished.

major comments (4)
  1. [III-D, Table II] The abstract's headline claim ('The late fusion of the best models outperforms the acoustic and text baselines by joint recognition balance metrics of 12.3% and 10.4%') is not supported as stated, because the comparison is between the late-fused SSL-trained models (best-2 fusion, test JRBM 0.430) and the two single-modality baselines (HuBERT 0.307, RoBERTa 0.326). The paper never reports a late fusion of the two baselines without SSL. Under the margin-sampling fusion rule, even two mediocre unimodal classifiers can jointly improve the harmonic-mean metric if their errors are complementary. Without a baseline-fusion control, the observed 12.3%/10.4% gains cannot be causally attributed to semi-supervised learning; they may be entirely due to modality fusion. Please add the fusion of the HuBERT baseline and RoBERTa baseline under the same late-fusion rule, and preferably also fusions of the SSL-trained unimodal models, before making the causal claim in the abstract.
  2. [III-D, Table II] All results are from a single run on a single re-split (449 test samples), with no error bars, multiple seeds, or statistical significance tests. This is particularly concerning because several SSL configurations underperform the corresponding baseline (e.g., fix-match with time masking: test JRBM 0.299 vs baseline 0.307; full-match with time masking: 0.296). The reported best improvement of 0.123 absolute JRBM may therefore be within run-to-run variance. Please report the mean and standard deviation over at least three seeds for the key comparisons (baselines, best SSL configurations, and fusions), or provide a significance test (e.g., a paired bootstrap or McNemar's test) for the test-set predictions.
  3. [II-B-2, Eqs. (2)-(4)] The equations for full-match learning appear internally inconsistent. Equation (2) and Eq. (4) apply the losses only when Rank(f(x_w^u)) > k, whereas the target y_e^u in Eq. (3) uses an indicator Rank(f(x_w^u)) ∈ [2,k]. For the samples on which Eq. (4) is active, y_e^u becomes 1, so the loss reduces to -log f(x_s^u), which maximizes the model's probability for the strongly augmented sample regardless of the pseudo label. This does not match the text's description of regulating the {2,...,k}-th classes. In addition, the text says 'The value of k is determined when the top-k accuracy is larger than a threshold σ,' which does not specify a search procedure or a per-sample versus global k. Please rewrite this subsection with a consistent notation, define the exact masks, and clarify how k is selected.
  4. [II-B-3] In the multi-task extension, the paper states that pseudo labels are used only when the model is confident in both emotion and intent predictions, but it does not provide the exact loss equation or how the two task-specific masks are combined (e.g., whether the threshold condition is applied to the product, the minimum, or the average of the two confidences). This affects the effective number of pseudo-labelled samples used for training and is necessary for reproducing the method. Please specify the combined loss explicitly.
minor comments (6)
  1. [II-B-2] The term 'entropy meaning loss' is unusual; 'entropy-minimizing loss' or 'entropy regularizer' would be clearer.
  2. [III-C] The statement 'All hyperparameters are set experimentally' is vague; please report the augmentation parameter values (e.g., flip duration, number of synonym substitutions) and the search ranges for τ, σ, and the loss coefficients.
  3. [III-D, Table II] The baseline rows are duplicated across the 'wo/ weak augmentation' and 'w/ weak augmentation' columns; displaying them once would avoid confusion.
  4. [III-A] Please state whether the re-split of the original training set uses a fixed random seed, and provide the seed or a reference to a script for reproducibility.
  5. [Abstract] 'joint recognition balance metrics' should be 'Joint Recognition Balance Metric (JRBM)', since Eq. (6) defines a single scalar.
  6. [IV] In the first sentence of the conclusions, 'Ultimately.' should be 'Ultimately,'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out evaluation and explicit SSL losses; headline comparison confound is an experimental-design issue, not circularity.

full rationale

The paper is an empirical study with no analytic derivation chain whose conclusions reduce to their inputs. The semi-supervised losses in Eqs. (1)-(5) use the model's own predictions on weakly augmented unlabelled data as pseudo-labels for strongly augmented versions, which is standard FixMatch-style consistency regularization rather than a definitional or fitted-input circularity; the test labels are never used to set these targets or to select hyperparameters. Model selection and hyperparameters are based on a re-split validation set, while the reported test JRBM comes from the original official validation set, so the headline result is not fitted to the test labels. Self-citations such as [7], [23], and [28] are ordinary method or task citations and are not load-bearing: the margin-sampling fusion rule is fully described in the text and does not presuppose the claimed improvement. The missing baseline-fusion control is a genuine experimental confound—the 12.3%/10.4% headline compares a fused SSL pair against single-modality baselines, so the gain cannot be cleanly attributed to SSL—but that is a validity/interpretation problem, not circularity. No equation, citation, or definition in the paper makes a claimed prediction equivalent by construction to a fitted input or to a self-cited uniqueness result, and therefore no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The ledger shows the central claim is empirical rather than derivation-based. It depends on distributional assumptions about unlabeled data, hand-chosen hyperparameters, and the reliability of Whisper transcriptions. No new physical or mathematical entities are introduced.

free parameters (4)
  • pseudo-label confidence threshold tau = 0.95
    Chosen in Section III-C; controls how many unlabeled samples contribute to the fix-match loss in Equation (1).
  • top-k accuracy threshold sigma = 0.99
    Chosen in Section III-C; determines k in the full-match adaptive negative and entropy-meaning losses.
  • loss coefficients lambda1, lambda2, lambda3 = 0.5
    Set equal in Section III-C; balance supervised and unsupervised terms in Equations (1) and (5).
  • augmentation hyperparameters = time mask up to 30k frames, pitch shift 4 steps, Gaussian noise scale 0.05, etc.
    Set experimentally in Section III-C; the choice of weak and strong augmentations changes whether SSL helps or hurts.
assumptions (3)
  • domain assumption Unlabeled MC-EIU samples come from the same distribution as labeled samples, and Whisper transcripts of unlabeled audio are accurate enough for pseudo-labeling the text model.
    Used in Section III-C to create text inputs for both fix-match and full-match; ASR errors are not measured or filtered.
  • domain assumption Thresholded pseudo-labels with tau=0.95 are reliable enough to provide a useful training signal.
    Core to Equation (1) in Section II-B1; if pseudo-labels are wrong, the added loss can degrade performance.
  • domain assumption A single re-split of MC-EIU into train, validation, and test estimates true performance on the challenge distribution.
    Section III-A; the official validation set is used as test, with no cross-validation or multiple seeds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-end Acoustic-linguistic Emotion and Intent Recognition Enhanced by Semi-supervised Learning." pith.science (2026). https://pith.science/paper/ROLMWFEJ

@misc{pith2026250707806,
  author       = {Pith},
  title        = {Pith review of: End-to-end Acoustic-linguistic Emotion and Intent Recognition Enhanced by Semi-supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROLMWFEJ}},
  note         = {Machine review of arXiv:2507.07806}
}
read the original abstract

Emotion and intent recognition from speech is essential and has been widely investigated in human-computer interaction. The rapid development of social media platforms, chatbots, and other technologies has led to a large volume of speech data streaming from users. Nevertheless, annotating such data manually is expensive, making it challenging to train machine learning models for recognition purposes. To this end, we propose applying semi-supervised learning to incorporate a large scale of unlabelled data alongside a relatively smaller set of labelled data. We train end-to-end acoustic and linguistic models, each employing multi-task learning for emotion and intent recognition. Two semi-supervised learning approaches, including fix-match learning and full-match learning, are compared. The experimental results demonstrate that the semi-supervised learning approaches improve model performance in speech emotion and intent recognition from both acoustic and text data. The late fusion of the best models outperforms the acoustic and text baselines by joint recognition balance metrics of 12.3% and 10.4%, respectively.

Figures

Figures reproduced from arXiv: 2507.07806 by the authors.

Figure 1
Figure 1. The pipeline of the semi-supervised learning method. The purple lines indicate loss functions in fix-match learning, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Confusion matrix of fusing the best 2 models for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Confusion matrix of fusing the best 2 models for [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 25 canonical work pages

  1. [1]

    Schuller and A

    B. Schuller and A. Batliner, Computational paralinguistics: Emotion, affect and personality in speech and language processing . John Wiley & Sons, 2013

    work page 2013

  2. [2]

    Domain adaptation in intent classification systems: A review,

    J. Atuhurra, H. Kamigaito, T. Watanabe, and E. Nichols, “Domain adaptation in intent classification systems: A review,” 2024, arXiv preprint:2404.14415

    arXiv 2024

  3. [3]

    End-to-end continuous speech emotion recognition in real-life customer service call center conversations,

    Y . Feng and L. Devillers, “End-to-end continuous speech emotion recognition in real-life customer service call center conversations,” in Proc. ACIIW, 2023, pp. 1–8

    work page 2023

  4. [4]

    A VEC 2019 workshop and challenge: State-of-mind, detecting depression with AI, and cross-cultural affect recognition,

    F. Ringeval, B. Schuller et al., “A VEC 2019 workshop and challenge: State-of-mind, detecting depression with AI, and cross-cultural affect recognition,” in Proc. AVEC, 2019, pp. 3–12

    work page 2019

  5. [5]

    Cross- lingual/cross-channel intent detection in contact-center conversations,

    S. Agrawal, A. Sachdeva, S. Jain, C. George, and J. Vepa, “Cross- lingual/cross-channel intent detection in contact-center conversations,” in Proc. INTERSPEECH, 2023, pp. 5269–5270

    work page 2023

  6. [6]

    Survey of deep representation learning for speech emotion recognition,

    S. Latif, R. Rana, S. Khalifa, R. Jurdak, J. Qadir, and B. Schuller, “Survey of deep representation learning for speech emotion recognition,” IEEE Transactions on Affective Computing , vol. 14, no. 2, pp. 1634– 1654, 2021

    work page 2021

  7. [7]

    Fast yet effective speech emotion recognition with self-distillation,

    Z. Ren, T. T. Nguyen, Y . Chang, and B. W. Schuller, “Fast yet effective speech emotion recognition with self-distillation,” in Proc. ICASSP, 2023, 5 pages

    work page 2023

  8. [8]

    An effective multimodal representation and fusion method for multimodal intent recognition,

    X. Huang, T. Ma, L. Jia, Y . Zhang, H. Rong, and N. Alnabhan, “An effective multimodal representation and fusion method for multimodal intent recognition,” Neurocomputing, vol. 548, p. 126373, 2023

    work page 2023

Show all 32 references
  1. [9]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems, vol. 33, pp. 12 449– 12 460, 2020

    2020

  2. [10]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 3451–3460, 2021

    2021

  3. [11]

    RoBERTa: A robustly optimized BERT pretraining approach,

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “RoBERTa: A robustly optimized BERT pretraining approach,” 2019, arXiv preprint:1907.11692

    2019 arXiv

  4. [12]

    Semi- supervised spoken language understanding via self-supervised speech and language model pretraining,

    C.-I. Lai, Y .-S. Chuang, H.-Y . Lee, S.-W. Li, and J. Glass, “Semi- supervised spoken language understanding via self-supervised speech and language model pretraining,” in Proc. ICASSP, 2021, pp. 7468– 7472

    2021

  5. [13]

    Semi-supervised speech emotion recognition with ladder networks,

    S. Parthasarathy and C. Busso, “Semi-supervised speech emotion recognition with ladder networks,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 28, pp. 2697–2709, 2020

    2020

  6. [14]

    Multi-task semi-supervised adversarial autoencoding for speech emo- tion recognition,

    S. Latif, R. Rana, S. Khalifa, R. Jurdak, J. Epps, and B. W. Schuller, “Multi-task semi-supervised adversarial autoencoding for speech emo- tion recognition,” IEEE Transactions on Affective Computing , vol. 13, no. 2, pp. 992–1004, 2020

    2020

  7. [15]

    Speech emotion recognition using semi-supervised learning with ladder networks,

    J. Huang, Y . Li, J. Tao, Z. Lian, M. Niu, and J. Yi, “Speech emotion recognition using semi-supervised learning with ladder networks,” in Proc. ACII Asia , 2018, 5 pages

    2018

  8. [16]

    Enhanced semi-supervised learning for multimodal emotion recognition,

    Z. Zhang, F. Ringeval, B. Dong, E. Coutinho, E. Marchi, and B. Schüller, “Enhanced semi-supervised learning for multimodal emotion recognition,” in Proc. ICASSP, 2016, pp. 5185–5189

    2016

  9. [17]

    Semi-fedSER: Semi-supervised learning for speech emotion recognition on federated learning using multiview pseudo-labeling,

    T. Feng and S. Narayanan, “Semi-fedSER: Semi-supervised learning for speech emotion recognition on federated learning using multiview pseudo-labeling,” in Proc. INTERSPEECH, 2022, pp. 5050–5054

    2022

  10. [18]

    Combining active and semi-supervised learning for spoken language understanding,

    G. Tur, D. Hakkani-Tür, and R. E. Schapire, “Combining active and semi-supervised learning for spoken language understanding,” Speech Communication, vol. 45, no. 2, pp. 171–186, 2005

    2005

  11. [19]

    Semi-supervised acoustic model training by discriminative data selection from multiple asr systems’ hypotheses,

    S. Li, Y . Akita, and T. Kawahara, “Semi-supervised acoustic model training by discriminative data selection from multiple asr systems’ hypotheses,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 24, no. 9, pp. 1524–1534, 2016

    2016

  12. [20]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,” Advances in neural information processing systems , vol. 33, pp. 596–608, 2020

    2020

  13. [21]

    Combining cross-modal knowledge transfer and semi-supervised learning for speech emotion recognition,

    S. Zhang, M. Chen, J. Chen, Y .-F. Li, Y . Wu, M. Li, and C. Zhu, “Combining cross-modal knowledge transfer and semi-supervised learning for speech emotion recognition,” Knowledge-Based Systems, vol. 229, p. 107340, 2021

    2021

  14. [22]

    Multi-lingual multi-task speech emotion recognition using wav2vec 2.0,

    M. Sharma, “Multi-lingual multi-task speech emotion recognition using wav2vec 2.0,” in Proc. ICASSP, 2022, pp. 6907–6911

    2022

  15. [23]

    Knowledge transfer for on-device speech emotion recognition with neural structured learning,

    Y . Chang, Z. Ren, T. T. Nguyen, K. Qian, and B. W. Schuller, “Knowledge transfer for on-device speech emotion recognition with neural structured learning,” in Proc. ICASSP, 2023, 5 pages

    2023

  16. [24]

    Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset,

    R. Liu, H. Zuo, Z. Lian, X. Xing, B. W. Schuller, and H. Li, “Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset,” 2024, arXiv preprint:2407.02751

    2024 arXiv

  17. [25]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” in Proc. ICML, 2023, 27 pages

    2023

  18. [26]

    Dynamic thresholding on fixmatch with weak and strong data augmentations for sound event detection,

    T. Khandelwal and R. K. Das, “Dynamic thresholding on fixmatch with weak and strong data augmentations for sound event detection,” in Proc. ISCSLP, 2022, pp. 428–432

    2022

  19. [27]

    EDA: Easy data augmentation techniques for boosting performance on text classification tasks,

    J. Wei and K. Zou, “EDA: Easy data augmentation techniques for boosting performance on text classification tasks,” in Proc. EMNLP- IJCNLP, 2019, 9 pages

    2019

  20. [28]

    Deep scalogram representations for acoustic scene classification,

    Z. Ren, K. Qian, Z. Zhang, V . Pandit, A. Baird, and B. Schuller, “Deep scalogram representations for acoustic scene classification,” IEEE/CAA Journal of Automatica Sinica , vol. 5, no. 3, pp. 662–669, 2018

    2018

  21. [29]

    Verification of acoustical correlates of emotional speech using formant-synthesis,

    F. Burkhardt and W. F. Sendlmeier, “Verification of acoustical correlates of emotional speech using formant-synthesis,” in Proc. ISCA ITRW , 2000, 6 pages

    2000

  22. [30]

    The acoustic realization of anger, fear, joy and sadness in chinese,

    J. Yuan, L. Shen, and F. Chen, “The acoustic realization of anger, fear, joy and sadness in chinese,” in Proc. ICSLP, 2002, 4 pages

    2002

  23. [31]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao et al. , “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  24. [32]

    Can large language models aid in annotating speech emotional data? Uncovering new frontiers,

    S. Latif, M. Usama, M. I. Malik, and B. W. Schuller, “Can large language models aid in annotating speech emotional data? Uncovering new frontiers,” 2023, arXiv preprint:2307.06090

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.